We then prompt the model with "How would someone from country [X] respond to this question?" Surprisingly, this makes model responses more similar to those of human respondents for some of the specified countries (i.e., China and Russia).

→ View original post on X — @anthropicai